Competing-risk neural networks in breast cancer: Generalization performance

New approaches to individualized therapy in breast cancer will require not only high performance in risk assessment for relapse, but also differential forecasting of the relative probability of competing risks. To achieve high performance even with factor interactions, a new formulation of competing-risk neural networks has been developed to determine the underlying risk associated with distinct relapse modes. Methods: A competing-risk neural network was trained on 2424 of 3424 primary breast cancer patients using previously published data from Munich and Rotterdam (38.5 % recurrence within follow-up) to produce differential risk scores for bone metastasis, distant metastasis other than bone, and loco-regional recurrence. Classical and tumor biological factors, country, as well as adjuvant systemic therapy variables were included. Differential risk scores were then obtained from the trained net for the remaining 1000 patients not previously seen by the net in order to test generalization performance, i.e., prediction and classification of disease recurrence. Results: The trained neural network is capable of generalization for all three recurrence categories considered. For example, grouping the neural bone scores into quartiles, the 5y bone-relapse-free survival rate in new patients is 56 % for the high-risk quartile vs. 88% for the rest. Conclusion: Statistical methods for survival capable of modeling factor interactions, such as neural networks, will become increasingly important in view of the prognostic information expected from new measurement techniques such as proteomics. However, probably the most important issue for acceptance of neural networks in survival modeling is generalization, i.e., the prognostic/predictive performance that can be reliably expected on new data. This paper illustrates that good generalization performance of competing risk neural networks in breast cancer can be attained using appropriate complexity reduction techniques. Hence, the present formulation of competing-risk neural networks is a promising candidate for improved risk assessment and ultimately for decision support in breast cancer.